Kandela

Persistent semantic memory for AI coding agents.
Your AI remembers decisions, failures, and context — across sessions, projects, and tools.

English | 한국어 | 日本語 | Deutsch | Français | Español | Português | 中文

▶ Watch demo on YouTube

---

Problem

Google, Claude, and ChatGPT each have their own memory — but it only works within their ecosystem. Switch from GPT to Claude, and all your memories are gone. You can't share context across tools or teammates.

Kandela solves this with cross-provider memory + full transparency + a 4-stage RAG pipeline.

Key Features

• 13 MCP Tools — Store, search, delete, update, auto-recall, on-demand search, inbox, project management
• Hybrid Search — Semantic + BM25 keyword search (RRF fusion)
• Importance Engine — Auto-scoring 1~10 + 18 rule-based infra tagging
• Lazy Retrieval — Brief mode (~260 tokens) + context_search on-demand
• Session Continuity — Detects environment changes (CWD, host, client) + auto-includes infra memories
• Prompt Guard — Prevents bad decisions based on stale memories
• Circuit Breaker — Detects repeated failure patterns + auto-stores gotchas
• Web Dashboard — Per-project memory browser, search, stats, performance monitoring
• One-click Install — Auto-installs hooks + 16 slash commands
• Multilingual Embeddings — paraphrase-multilingual-MiniLM-L12-v2 (50+ languages)

Benchmark

HIPAA medical data pipeline scenario (8 sessions, 14 decision traps) — Kandela ON vs OFF:

	Kandela ON	Kandela OFF	Delta
Decision Trap Avoidance	100%	11.9%	+88.1pp
Task Time	77.9 min	86.6 min	-10.1%
Generated Code	2,152 lines	3,441 lines	-37.5%
Generated Files	40	62	-35.5%

3 runs (seeds=42,123,456), claude-sonnet-4-6, Groq Llama 3.3 70B (Operator).

Key Insights

• Decisions not in code are what matters: Auditor names, OOM incidents, data loss history — information invisible to code reading
• Code-based decisions are self-defended by LLM: In code-centric scenarios (InfraBot), LLM achieved 95.2% avoidance even without Kandela
• Eliminates rework: Without Kandela, AI re-implements previously rejected approaches — 37.5% code waste

Large-Scale Data Analysis

830K real conversations (WildChat-1M) and 1M LLM conversations (LMSYS-Chat-1M):

Finding	Metric	Implication
Initial instruction reference drops 18%p after 10 turns	66%→48%	General LLMs start forgetting early context
Small LLMs drop up to 66%p	chatglm 84%→19%	Memory augmentation critical for local/small models
14% of user corrections are "forgot previous conversation"	82K coding conversations	Core problem Kandela solves
Kandela maintains context after 49+ turns	Internal benchmark	Brief Recall prevents decay

Architecture

Client (Claude Code / Desktop / Cursor / Telegram Bot)
  │
  ├─ MCP Protocol ──→ Kandela Server
  │                      │
  │                  MemoryStore (ChromaDB)
  │                      ├── Hybrid Search (Semantic + BM25 RRF)
  │                      ├── Importance Engine (18 rules)
  │                      ├── Session Continuity
  │                      └── Prompt Guard / Circuit Breaker
  │
  ├─ Telegram Bot ──→ Natural language memory access (LLM intent classification)
  │
  └─ Web Dashboard ──→ REST API + Memory Browser

Memory Types

Type	Description	Auto/Manual
fact	Preferences, tech stack, environment info	Both
decision	Design decisions, trade-offs, rationale	Manual
summary	Session summaries at conversation end	Auto
snippet	Frequently used code patterns, configs	Manual

Search Pipeline

Query → Semantic (MiniLM-L12) ──────┐
                                     ├─→ RRF Fusion → Importance Rerank → MMR Diversity → Results
Query → BM25 (multilingual NLP) ────┘
        ├── Korean: kiwipiepy morphological analysis
        └── Others: regex tokenizer (English, German, Spanish, etc.)

Get Started

Hosted Service (Recommended)

Request beta access, then install in 2 minutes.

# 1. Email support@kandela.ai for beta invite code
# 2. Sign up at https://api.kandela.ai/dashboard with the invite code
# 3. Generate API key in Account page
# 4. One-line install (prompts for API key, sets up everything):
curl -fsSL https://api.kandela.ai/api/install | bash

→ Getting Started Guide

Self-Hosted

Run your own Kandela instance. Single-user mode, full control over your data.

git clone https://github.com/deep-on/kandela-selfhost.git && cd kandela-selfhost/docker
docker compose up -d
# → http://localhost:8321/dashboard

→ Self-Host Repository

Hosted vs Self-Hosted

	Hosted (api.kandela.ai)	Self-Hosted
Setup	2 min	5 min
Multi-user	✅	Single-user
Telegram Bot	✅	—
Remote Commands	✅	—
Activity Heatmap	✅	—
Tier Features (Pro/Max)	✅	All features
Data Location	Cloud	Your server
Maintenance	Managed	Self-managed

Tech Stack

• Python 3.11+, FastMCP (mcp[cli])
• ChromaDB — vector database (persistent)
• sentence-transformers — local embeddings (paraphrase-multilingual-MiniLM-L12-v2)
• Pydantic v2 — input validation
• Docker — deployment

Links

• Homepage: kandela.ai
• Self-Host Repo: deep-on/kandela-selfhost
• Privacy Policy: kandela.ai/privacy
• Terms of Service: kandela.ai/terms

License

• Server: AGPL-3.0 — Copyright (c) 2025-2026 Deep-ON Inc.
• Client-side files (hooks, slash commands): MIT

Disclaimer

This software is provided "AS IS" without warranty of any kind. Users are responsible for backing up their own data.